MojiDiff

← experiments

masked-continuity-l12-start-full-7d18a83-2681icons-9b9b1699

completed —

Measured behaviour

Held-out negative log likelihood per free token
modelposition-marginal floor
nats per tokenoptimizer step02460200040006000selected
Table view
optimizer stepmodelposition-marginal floor
05.18564.7865
3004.39294.7865
6004.29064.7865
9004.22834.7865
12004.18344.7865
15004.15804.7865
18004.14124.7865
21004.11944.7865
24004.08634.7865
27004.06114.7865
30003.98554.7865
33003.89814.7865
36003.85214.7865
39003.79284.7865
42003.76614.7865
45003.75594.7865
48003.73454.7865
51003.71924.7865
54003.68314.7865
57003.70684.7865
60003.68444.7865
63003.68244.7865
Masked-token accuracy over unforced positions
modelposition-marginal floor
accuracyoptimizer step00.050.10.150.20200040006000selected
Table view
optimizer stepmodelposition-marginal floor
00.04790.1307
3000.13180.1307
6000.14690.1307
9000.14630.1307
12000.15060.1307
15000.15380.1307
18000.16680.1307
21000.16620.1307
24000.16250.1307
27000.16410.1307
30000.16680.1307
33000.16780.1307
36000.17160.1307
39000.17430.1307
42000.17270.1307
45000.17590.1307
48000.18560.1307
51000.18340.1307
54000.18500.1307
57000.19040.1307
60000.18990.1307
63000.18830.1307

Measured result

Verbatim from the run's summary.json.

checkpoint_round_trip
true
checkpoint_sha256
f48dd00addd4f612…
checks
all_valid
true
beats_copy_previous
false
config_sha256
73add6762374a1c1…
continuity
copy_previous
mean_bins_off
25.64749262536873
median_bins_off
17.0
n
339
marginal
mean_bins_off
140.05162241887905
median_bins_off
138.5
n
339
model
mean_bins_off
29.13864306784661
median_bins_off
22.0
n
339
coordinate_tau
1.0
criteria
beats_copy_previous
true
require_all_valid
true
deterministic_algorithms
true
device
cuda
evaluation_icons
339
evaluation_split
validation
first_train_loss
5.011204242706299
gpu
NVIDIA GeForce RTX 4080
held_out_nll_per_masked_token
3.682448462557575
icon_presentations
100800
initial_held_out_nll
5.185607692642582
inpainting
all_valid
true
attempted
16
close_reproduction_rate
0.0
decoding
chain_order
false
greedy_for_criteria
true
iterations
8
samples
1
drop
median_rgba_mae
0.009685249485596707
exact_reproduction_rate
0.0
icons
16
marginal
helped
2
mean_recovery
-29608927.159026496
median_recovery
-1.3892468653280403
median_rgba_mae
0.026510208635923502
model
helped
1
mean_recovery
-186376636.92392704
median_recovery
-0.7598014884600544
median_rgba_mae
0.01443723901597676
paired
drop_minus_model
interval
-0.0141243645042416, 0.006982598171937068
interval_excludes_zero_above
false
mean
-0.003570883166152266
n
16
positive
1
marginal_minus_model
interval
-0.0037150367756598993, 0.010215571982026397
interval_excludes_zero_above
false
mean
0.003250267603183249
n
16
positive
11
rows_sha256
9abec9fb93470474…
sheet_sha256
0b7961b9bcf64563…
last_train_loss
3.5597805976867676
loss_reduction_factor
1.4081955919733404
marginal_masked_token_accuracy
0.1307154384077461
marginal_nll_per_masked_token
4.786513732286344
mask_mixture
drawn
span
100800
max_paths
2
random_rate
0.05, 0.95
span_max
1
weights
span
1.0
masked_token_accuracy
0.18827326519634213
metric_head
false
metrics_sha256
e8b53454c35c9aa9…
model_parameters
530146
nll_ratio_to_marginal
0.7693383260803063
path_binding
false
predeclared_outcome
falsified
schema_version
1
selected_step
6300
sequence_length
1376
start_features
true
steps_run
6300
study_version
masked-continuity-l12-start-full
timing
peak_vram_gib
1.945967674255371
prepare_seconds
101.86072754301131
train_seconds
409.2479177750065
torch_version
2.14.0a0+4fdf77b940.nv26.08
train_icons
2681
vocabulary
418

Visual output

inpainting
inpainting.png

State transitions

  1. planned2026-09-21T09:32:44Z
  2. completed2026-09-21T09:47:45Zgiven the full budget, start features alone bring the continuity error to 29.1 bins (median 22.0) against the copy policy's 25.6 (median 17.0), held-out likelihood to 3.682 - 0.769 of the floor - and the curve is still falling at step 6,300, where the checkpoint was selected; short of the criterion by 3.5 bins on the mean

Run record

This run has no run.yaml. What follows is the identity and configuration carried by its rows in state/runs.jsonl, the append-only registry.

run_id
masked-continuity-l12-start-full-7d18a83-2681icons-9b9b1699
config
configs/learning/masked-continuity-l12-start-full.yaml
outputs
checkpoint_root
data/processed/masked-continuity-l12-start-full
report_root
reports/learning/masked-continuity-l12-start-full